qbm4eo.rbm module

This file is a part of qbm4eo.rst project.

It has been modified as a part of the EuroHPC PL project funded at the Smart Growth Operational Programme 2014-2020, Measure 4.2 under the grant agreement no. POIR.04.02.00-00-D014/20-00.

class qbm4eo.rbm.AnnealingRBMTrainer(num_steps: int, sampler: Sampler, qubo_scale: float = 1.0, learning_rate: float = 0.01, **sampler_kwargs: Dict[str, Any])[source]

Bases: RBMTrainer

A class for training the RBM using the annealing samplers (D-Wave of CIMSampler).

A default constructor for the AnnealingRBMTrainer class instances.

Parameters:

num_steps – Number of training steps.
sampler – A sampler to be used during the training.
qubo_scale – A scaling factor for the QUBO.
learning_rate – A learning rate.
sampler_kwargs – Additional keyword arguments to be passed to the sampler.

training_step(rbm: RBM, batch: ndarray[Any, dtype[float32]]) → None[source]

A single training step.

Parameters:

rbm – An RBM to be trained.
batch – The RBMs’ visible layer neurons’ states.

class qbm4eo.rbm.CD1Trainer(num_steps: int, learning_rate: float = 0.01)[source]

Bases: RBMTrainer

A class for training the RBM using the Contrastive Divergence 1 algorithm.

A default constructor for the CD1Trainer class instances.

Parameters:

num_steps – Number of training steps.
learning_rate – A learning rate.

training_step(rbm: RBM, batch: ndarray[Any, dtype[float32]]) → None[source]

A single training step.

Parameters:

rbm – An RBM to be trained.
batch – The RBMs’ visible layer neurons’ states. TODO: There seems to be a problem when batch_size is 1. Investigate.

class qbm4eo.rbm.RBM(num_visible: int, num_hidden: int, rng: Generator | None = None)[source]

Bases: object

A class implementing the Restricted Boltzmann Machine.

A default constructor for the RMB class instances.

Parameters:

num_visible – Number of neurons in the visible layer.
num_hidden – Number of neurons in the hidden layer.
rng – Random number generator to be used by the RBM.

h_probs_given_v(v_batch: ndarray[Any, dtype[float32]]) → ndarray[Any, dtype[float32]][source]

Compute the probabilities of the hidden layer neurons being active given the visible layer neurons’ states.

Parameters:: v_batch – A batch of visible layer neurons’ states.
Returns:: A batch of probabilities of the hidden layer neurons being active.

classmethod load(file: str | BytesIO) → RBM[source]

Load the RBM from a file.

Parameters:: file – A file to load the RBM from.
Returns:: A loaded RBM.

reconstruct(v_batch: ndarray[Any, dtype[float32]]) → ndarray[Any, dtype[float32]][source]

Reconstruct the visible layer neurons’ states given the visible layer neurons’ states.

Parameters:: v_batch – A batch of visible layer neurons’ states.
Returns:: A batch of reconstructed visible layer neurons’ states.

sample_h_given_v(v_batch: ndarray[Any, dtype[float32]]) → ndarray[Any, dtype[float32]][source]

Sample the hidden layer neurons’ states given the visible layer neurons’ states.

Parameters:: v_batch – A batch of visible layer neurons’ states.
Returns:: A batch of hidden layer neurons’ states.

sample_v_given_h(h_batch: ndarray[Any, dtype[float32]]) → ndarray[Any, dtype[float32]][source]

Sample the visible layer neurons’ states given the hidden layer neurons’ states.

Parameters:: h_batch – A batch of hidden layer neurons’ states.
Returns:: A batch of visible layer neurons’ states.

save(file: str | BytesIO) → None[source]

Save the RBM to a file.

Parameters:: file – A file to save the RBM to.

v_probs_given_h(h_batch: ndarray[Any, dtype[float32]]) → ndarray[Any, dtype[float32]][source]

Compute the probabilities of the visible layer neurons’ states given the hidden layer neurons’ states.

Parameters:: h_batch – A batch of hidden layer neurons’ states.
Returns:: A batch of probabilities of the visible layer neurons’ states.

class qbm4eo.rbm.RBMTrainer(num_steps: int)[source]

Bases: object

A base class for implementing the RBM training algorithms.

TODO TR: There are some nasty casting between np.matrix and np.ndarray that: is there to ensure proper transposition or multiplications when one of the matrix dimension is 1. This probable could be done more elegantly.

A default constructor for the RBMTrainer class instances.

Parameters:: num_steps – Number of training steps.

fit(rbm: RBM, data_loader: DataLoader[Tuple[Any, Any]] | Generator[Tuple[Any, Any], Any, Any], callback: Callable[[int, RBM, float], None] = None, verbose: bool = False) → None[source]

Fits the RBM to the data.

Parameters:

rbm – An RBM to be trained.
data_loader – A data loader.
callback – A callback function to be called after each training step.

abstractmethod training_step(rbm: RBM, batch: ndarray[Any, dtype[float32]]) → None[source]

A single training step.

Parameters:

rbm – An RBM to be trained.
batch – A batch of data.

qbm4eo.rbm.infinite_dataloader_generator(data_loader: DataLoader[Tuple[Any, Any]] | Generator[Tuple[Any, Any], Any, Any]) → Generator[Tuple[Any, Any], None, None][source]

A generator that infinitely yields data batches from the given dataloader.

Parameters:: data_loader – A PyTorch DataLoader initialized with the desired dataset.
Returns:: Yields an index of a batch and the data from the data_loader.

qbm4eo.rbm.qubo_from_rbm_coefficients(weights: Tensor | ndarray[Any, dtype[float32]], v_bias: Tensor | ndarray[Any, dtype[float32]], h_bias: Tensor | ndarray[Any, dtype[float32]]) → BinaryQuadraticModel[source]

Create a QUBO problem representing RBM with given coefficients.

Parameters:

weights – A square interaction matrix.
v_bias – An N-element visible layer bias vector.
h_bias – An M-element hidden layer bias vector.

Returns:

A QUBO, represented as dimod.BQM with N+M variables, such that:

variables 0,…,N-1 correspond to hidden layer
variables 0,…,M-1 correspond to visible layer
visible layer biases correspond to linear coefficients of first N variables
hidden layer biases correspond to linear coefficients of second M variables
weights correspond to interaction terms between first N and second M variables.

Note

This function does not allow for manipulating how RBM variables are mapped to the QUBO variables. This is not a problem if QUBO is to be used with an unstructured sampler. For sampler, the intended usage is to wrap it in the EmbeddingComposite.